Abstract Conditional Logistic GLMM Inference

Fits one likelihood with a conditional-logistic contribution from discordant matched pairs and a random-intercept logistic GLMM contribution from concordant matched pairs and reservoir subjects.

Known estimability limitation at high covariate counts. The treatment effect is identified from the discordant-pair conditional-logit term alone (the shared pair-level intercept cancels only for pairs that disagree on outcome), so that term alone must identify treatment plus every adjustment covariate from however many pairs happen to be discordant for a given incidence response, typically a modest fraction of \(n\). With model_formula = ~. on a dataset with many covariates, the number of free parameters the discordant term must help identify can approach or exceed the number of discordant pairs, and the joint fit then frequently converges to an out-of-bounds coefficient or random-effect log-variance and is correctly reported as nonestimable rather than returned. This was measured directly (comprehensive_tests.R repro, 2026-10): the identical design/response/n was estimable in 39 of 40 null replicates with one covariate, but only 22 of 40 with ten covariates under model_formula = ~.. This is treated as an accepted property of this estimator in this regime, not a defect to be patched; when the cause is specifically this parameter-to-discordant-pair imbalance, the cached nonestimability reason names it explicitly (joint_likelihood_nonestimable_high_dimensional_relative_to_discordant_pairs, with the actual counts) rather than the generic joint_likelihood_nonestimable that covers every other cause of the same two bound checks.

Super class

Inference -> InferenceAbstractKKCondLogitGLMM

Methods

Public methods

• InferenceAbstractKKCondLogitGLMM$new()

• InferenceAbstractKKCondLogitGLMM$compute_estimate()

• InferenceAbstractKKCondLogitGLMM$compute_estimate_with_bootstrap_weights()

• InferenceAbstractKKCondLogitGLMM$compute_asymp_confidence_interval()

• InferenceAbstractKKCondLogitGLMM$compute_asymp_two_sided_pval()

• InferenceAbstractKKCondLogitGLMM$clone()

+ inherited public methods from Inference

• Inference$capabilities()
• Inference$compute_exact_confidence_interval()
• Inference$compute_exact_two_sided_pval_for_treatment_effect()
• Inference$duplicate()
• Inference$get_analysis_data()
• Inference$get_covariates()
• Inference$get_design_object()
• Inference$get_model_formula()
• Inference$get_nonestimable_reason()
• Inference$get_nonestimable_stage()
• Inference$get_optimization_alg()
• Inference$get_response()
• Inference$get_response_type()
• Inference$get_treatment()
• Inference$is_nonestimable()
• Inference$set_optimization_alg()
• Inference$set_seed()
• Inference$supports()

---

InferenceAbstractKKCondLogitGLMM$new()

Initialize KK conditional-logit GLMM incidence inference, validate the binary response, and prepare the matched-pair conditional likelihood and reservoir mixed-model components. See InferenceAbstractKKCondLogitGLMM.

Usage

InferenceAbstractKKCondLogitGLMM$new(
  des_obj,
  model_formula = NULL,
  max_abs_reasonable_coef = 50,
  max_abs_reasonable_se = 10,
  max_abs_log_sigma = 8,
  verbose = FALSE,
  smart_cold_start_default = NULL,
  optimization_alg = NULL
)

Arguments

des_obj

A completed Design object with an incidence or proportion response.

model_formula

Optional formula for covariate adjustment.

max_abs_reasonable_coef

Cap for reasonable coefficient estimates.

max_abs_reasonable_se

Cap for reasonable treatment standard errors.

max_abs_log_sigma

Cap for reasonable log random effect variance.

verbose

Logical. Whether to print progress messages.

smart_cold_start_default

Logical. Whether to use smart starting values for the optimizer.

optimization_alg

Character. Optimization algorithm (default "lbfgs").

---

InferenceAbstractKKCondLogitGLMM$compute_estimate()

Computes the class-specific treatment-effect estimate; see Inference.

Usage

InferenceAbstractKKCondLogitGLMM$compute_estimate(estimate_only = FALSE)

Arguments

estimate_only

Logical. If TRUE, skip variance component calculations.

---

InferenceAbstractKKCondLogitGLMM$compute_estimate_with_bootstrap_weights()

Recomputes the class-specific treatment estimate for a bootstrap sample; see InferenceNonParamBootstrap.

Usage

InferenceAbstractKKCondLogitGLMM$compute_estimate_with_bootstrap_weights(
  subject_or_block_weights,
  estimate_only = FALSE
)

Arguments

subject_or_block_weights

Numeric vector. Row weights for bootstrap.

estimate_only

Logical. If TRUE, skip variance component calculations.

---

InferenceAbstractKKCondLogitGLMM$compute_asymp_confidence_interval()

Uses the shared asymptotic confidence-interval contract; see InferenceAsymp.

Usage

InferenceAbstractKKCondLogitGLMM$compute_asymp_confidence_interval(
  alpha = 0.05
)

Arguments

alpha

Numeric. Significance level (default 0.05).

---

InferenceAbstractKKCondLogitGLMM$compute_asymp_two_sided_pval()

Uses the shared asymptotic two-sided p-value contract; see InferenceAsymp.

Usage

InferenceAbstractKKCondLogitGLMM$compute_asymp_two_sided_pval(delta = 0)

Arguments

delta

Numeric. Null treatment effect value (default 0).

---

InferenceAbstractKKCondLogitGLMM$clone()

The objects of this class are cloneable with this method.

Usage

InferenceAbstractKKCondLogitGLMM$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.